Improved YOLOv11n-Based Detection Algorithm for Railway Surface Defect Inspection

Abstract

Railway surface defect detection is a critical component of automated track maintenance and safety assurance. Standard deep learning detectors face three domain-specific challenges in rail inspection: high visual similarity between de-fects and metallic rail surface texture, large spatial scale disparity across defect categories, and insufficient receptive field diversity for capturing large-area peri-odic corrugation patterns. To address these challenges, this paper proposes RSD-YOLO, an improved YOLOv11n-based detector incorporating three targeted structural modifications. First, Convolutional Block Attention Modules (CBAM) are inserted after the P3 and P4 backbone stages to provide multi-level dual-axis feature recalibration, suppressing metallic background noise and amplifying de-fect-discriminative activations. Second, BiFPN adaptive weighted fusion (BiFPN_Concat2) replaces standard PANet concatenation in the neck, enabling learnable asymmetric cross-scale feature interaction. Third, Atrous Spatial Pyra-mid Pooling (ASPP) with dilation rates {6, 12, 18} replaces the SPPF module at the backbone apex to provide parallel multi-receptive-field context aggregation without resolution loss. Experiments on a custom 8,000-image rail defect dataset of four categories demonstrate that RSD-YOLO achieves [email protected] of 84.7% and [email protected]:0.95 of 64.8%, outperforming the YOLOv11n baseline by 3.8% and 3.4% respectively. Wheel Burn detection improves by 6.8% and Corrugation by 5.6%, confirming the model's effectiveness for the most diagnostically challenging defect categories in railway surface inspection.